Hello, excuse me. Are this images in the LIP dataset cropped based on the bounding boxes annotated in MSCOCO, or you just annotate in manual?